%3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>

    %16 = "tf.Cast"(%15) {Truncate = false} : (tensor<2x2x2xi32>) -> tensor<2x2x2xf32>
    %17 = "tf.Mul"(%16, %cst_0) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %18 = "tf.AddV2"(%17, %cst) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %19 = "tf.Floor"(%18) : (tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
    %20 = "tf.ClipByValue"(%19, %cst_5, %cst_6) : (tensor<2x2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2x2xf32>

func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}

module {
func.func @tfInplaceUpdate(%arg0: tensor<2x1x2xf32>) -> tensor<2x1x2xf32> {
  %1 = arith.constant dense<1> : tensor<1xi32>
  %2 = arith.constant dense<2.0> : tensor<1x1x2xf32>
  %3 = "tf.InplaceUpdate"(%arg0, %1, %2) {device = ""}
    : (tensor<2x1x2xf32>, tensor<1xi32>, tensor<1x1x2xf32>) -> tensor<2x1x2xf32>
  func.return %3 : tensor<2x1x2xf32>
}
}

//CHECK: module attributes

  func.return %17 : tensor<1x2x3xf32>

  // CHECK: %[[NONE:.*]] = "tfl.no_value"() <{value}> : () -> none
  // CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

  
  %40 = "tfl.read_variable"(%4) : (tensor<!tf_type.resource>) -> tensor<1x2x3xf32>
  %41 = "quantfork.stats"(%40) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %42 = "tfl.concatenation"(%41, %0) {axis = 1 : i32, fused_activation_function = "NONE"} : (tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x4x3xf32>

// CHECK-NEXT:    return %2 : tensor<3x2xf32>
}

func.func @transpose_3d(%arg0: tensor<1x2x3xf32>) -> tensor<3x2x1xf32> {
  %0 = "mhlo.transpose"(%arg0) <{permutation = dense<[2, 1, 0]> : tensor<3xi64>}> : (tensor<1x2x3xf32>) -> tensor<3x2x1xf32>
  func.return %0 : tensor<3x2x1xf32>

// CHECK-LABEL:   transpose_3d

        none, none, none, none) -> tensor<1x2x3xf32>
  %17 = "quantfork.stats"(%16) {
    layerStats = dense<[0.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],
      [-8.0, 8.0],
      [-0.5, 0.5]
    ]> : tensor<3x2xf32>, axis = 2 : i64
  } : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  func.return %17 : tensor<1x2x3xf32>

    %6 = "quantfork.qcast"(%5) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>
    %7 = "quantfork.dcast"(%6) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>) -> tensor<1x3x2x2xf32>
    return %7 : tensor<1x3x2x2xf32>
  }

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";